LATE G. N. SAPKAL COLLEGE OF ENGINEERING

Exploring Deep Neural Architectures for Skin Lesion Detection and Classification

Skin cancer, such as melanoma, is an important global public health issue, in which the early detection and appropriate classification contributed to establish an early diagnosis that resulted in a more effective and efficient treatment. Recent advances in deep learning have sparked a new revolution in medical image analysis that has led to powerful automatic methods for skin lesion detection and classification. In this paper, we discuss the design and application of deep neural architectures, i.e., Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), Recurrent Neural Networks (RNNs), and hybrid models, for the purpose of dermatology diagnostics. We also provide a comparative study of benchmark datasets (ISIC, HAM10000, and PH2), efficacy of the preprocessing methods, and data augmentation and transfer learning strategies as a means of inflating performance metrics. The critical issues such as class imbalance, interpretability and cross-dataset generalization are also comprehensively addressed. This work also suggests some future research directions, including explainable AI and multimodal learning for dermatology. The objective of the review is to provide the researchers and clinicians in general, the overall perception of pros and cons of the available methods and the inspiration for creativity towards smart diagnostic systems of detecting and classifying skin cancer.